Advancements in neural network techniques for electric and magnetic field reconstruction: Application to ion radiography

• Published in: AIP advances 2024-02, Vol.14 (2), p.025037-025037-12
• Main Authors: Jao, C.-S., Chen, Y.-C., Nikaido, F., Liu, Y.-L., Sakai, K., Minami, T., Isayama, S., Abe, Y., Kuramitsu, Y.
• Format: Article
• Language: English
• Subjects: Charged particles; Correlation coefficients; Electromagnetic fields; Electromagnetism; Exact solutions; Image reconstruction; Magnetic fields; Mathematical analysis; Neural networks; Plasma; Radiography; Training
• ISSN: 2158-3226; 2158-3226
• DOI: 10.1063/5.0189878

In the realm of high-energy-density laboratory plasma experiments, ion radiography is a vital tool for measuring electromagnetic fields. Leveraging the deflection of injected protons, ion imaging can reveal the intricate patterns of electromagnetic fields within the plasma. However, the complex task of reconstructing electromagnetic fields within the plasma system from ion images presents a formidable challenge. In response, we propose the application of neural network techniques to facilitate electromagnetic field reconstructions. For the training data, we generate corresponding particle data on ion radiography with diverse field profiles in the plasma system, drawing from analytical solutions of charged particle motions and test-particle simulations. With these training data, our expectation is that the developed neural network can assimilate information from ion radiography and accurately predict the corresponding field profiles. In this study, our primary emphasis is on developing these techniques within the context of the simplest setups, specifically uniform (single-layer) or two-layer systems. We begin by examining systems with only electric or magnetic fields and subsequently extend our exploration to systems with combined electromagnetic fields. Our findings demonstrate the viability of employing neural networks for electromagnetic field reconstructions. In all the presented scenarios, the correlation coefficients between the actual and neural network-predicted values consistently reach 0.99. We have also learned that physics concepts can help us understand the weaknesses in neural network performance and identify directions for improvement.